The paper introduces Actionbook, an open-source solution designed to significantly improve the efficiency, speed, and reliability of AI agents performing web browsing and automation tasks. It addresses critical limitations of current agent frameworks, which typically process the entire Document Object Model (DOM) of a webpage.

Problem Statement:
Traditional web browsing agents feed the complete HTML DOM to Large Language Models (LLMs). This approach incurs several major drawbacks:

1. Excessive Token Consumption: Parsing an entire webpage's DOM can consume tens of thousands of tokens (e.g., for an Airbnb search), leading to high operational costs and quickly exhausting LLM context windows (e.g., over 60% of GPT-5's context for a single page).
2. Performance Degradation: The LLM struggles to identify relevant elements (like buttons) within a massive and unstructured DOM, akin to "groping in the dark," resulting in slow execution.
3. Fragile Automation: Agents often rely on hardcoded selectors that break when website UIs change, necessitating extensive agent logic modifications.
4. Increased Hallucination/Errors: Complex DOM structures can cause LLMs to generate incorrect actions or "hallucinate" due to misinterpretations.

Core Methodology - Actionbook's "Behavior Manual":
Actionbook, built upon Vercel's agent-browser, fundamentally shifts the paradigm from exploratory DOM parsing to pre-defined, structured knowledge. Its core methodology revolves around creating and utilizing a "behavior manual" for each specific website.

5. Pre-defined Action Mapping: For commonly automated websites, a manual is created that maps specific actions (e.g., "search for a flight," "click login") to their corresponding DOM selectors and interaction patterns.
6. Compressed JSON Representation: This "behavior manual" is then compressed into a compact JSON format. This structured representation contains only the essential information required for interaction, such as element IDs, class names, XPath, or CSS selectors for target elements, and the sequence of operations.
7. Context Injection: Instead of the voluminous raw HTML, this compact JSON manual is injected into the LLM's context. This dramatically reduces the input token count.
8. Direct Execution: The LLM, informed by this precise and concise manual, can then directly infer and execute the required actions without needing to process and understand the entire visual layout or reconstruct the interactive elements from scratch. The agent bypasses the exploratory phase of identifying elements, leading to immediate action.

Technical Benefits and Implementation:

• Token Efficiency: Reduces token usage by approximately 100 times, as the compressed JSON is significantly smaller than the full HTML DOM ($T_{JSON} \approx \frac{1}{100} T_{HTML}$).
• Speed Improvement: Achieves up to a 10x speed increase due to direct action execution and reduced LLM processing load.
• Enhanced Maintainability: UI changes on a website only require updating the site-specific JSON manual, not the LLM's core prompt or agent logic, thereby maintaining agent robustness.
• LLM Agnostic: The method is compatible with various LLMs (e.g., GPT-5.3-Codex, Claude Opus 4.6, Gemini 3 Pro).
• Version Control: The manuals can be version-controlled, ensuring greater stability for automated tasks.
• Optimized Implementation (actionbook-rs): A Rust-based implementation, actionbook-rs, is recommended over the TypeScript version due to superior performance characteristics:
  • Binary Size: 7.8 MB (Rust) vs. 150 MB (TypeScript).
  • Startup Time: Approximately 5 ms (Rust) vs. 500-800 ms (TypeScript).
  • Dependencies: Zero runtime dependencies for the Rust version, facilitating easier integration into CI/CD pipelines.
  • Browser Re-use: Utilizes existing Chrome/Brave installations, eliminating the need for separate browser installations.
  • Built-in Features: Includes stealth mode and cookie management.

Impact on Agent Consistency:
Integrating these pre-defined manuals as "skills" within coding agents (e.g., Claude Code) leads to significantly higher success rates and consistency for repetitive web automation tasks. This method is particularly effective for stable, repetitive web browsing automation rather than general development testing, for which tools like Playwright or Chrome DevTools are still recommended.